A real engineering changelog. We tag entries — NEW for genuine new features, IMPROVED for behaviour changes, FIXED for bugs, DEPRECATED for removals — and date them to the day they hit production.
All managed Postgres clusters can be upgraded to 17.2 with a single click in the console or arv pg upgrade --version 17.2. Logical replication lag during cutover stayed under 8 ms, p99.
V8 isolate spin-up dropped from 6.1 ms to 3.6 ms p99 on Ampere Altra hosts after we landed an upstream snapshot-warming patch. No action needed.
Bring your own IPv6 prefix, sign an RPKI ROA, and announce from any region group. Same flow as IPv4 BYOIP we shipped last December.
$ arv net byoip add 2a0c:5240::/48 --asn 64500 --roa-file roa.signedPre-signed URLs created within 200 ms of the X-Amz-Date floor occasionally returned 403 from the edge cache instead of the origin. Cache now reads the wall clock from the regional NTP rather than the PoP local clock.
Our second North American region opened on Q1's last day. Compute, Storage, and Functions are GA; Postgres and Kubernetes follow next month.
The web console now responds to all the keyboard shortcuts you'd expect: g d for dashboard, g c for compute, / to focus search, ? to see them all.
Copy-on-write branches from any LSN, available on every plan. Branches stay free of charge until they diverge — only the diverged pages are billed at €0.09/GB/month.
We finished the BFD + GoBGP rollout. The 99th percentile time from a region withdrawal to traffic settling on the next-best PoP is now 196 ms, down from 480 ms.
A bad BFD-multipath config caused a 14-minute partial outage in Frankfurt. Read the postmortem. SLA credits were applied automatically to affected accounts.
Multi-region VPCs now connect via a host-offloaded WireGuard mesh. Inter-region throughput is line-rate on a8 plans and up; latency is whatever physics allows (~23 ms hel↔fra).
Available immediately on new clusters. Existing clusters can upgrade with arv k8s upgrade --version 1.32; Karpenter will roll node pools without disruption to PDB-respecting workloads.
The original v0 REST API will return 410 Gone on June 30. Migrate to v1 — the spec is fully backward-compatible aside from the timestamp format (RFC 3339 instead of Unix ms).
Compute snapshots now use ZFS incremental send. A 320 GB NVMe snapshot used to average 38 s; it's now 4.2 s, p50.
Our first South American region. IX.br peering. 78% renewable power as of January's grid mix.
Available on all managed Postgres clusters. We've benchmarked recall@10 of 0.987 on a 12M-vector dataset at 4 ms p99 query latency on a p4 instance.
Customers outside the EU (notably AU, BR, JP) saw VAT lines in their PDF invoices despite reverse-charge applying. Fixed retroactively for invoices issued in November and December.
Bring your own /24, sign an RPKI ROA, attach to one or more regions. We announce from AS207443 with the BGP communities of your choice.
Per-PoP replicated KV with linearizable reads at the local PoP, eventual elsewhere. Replication lag p99 under 240 ms across 14 regions.
We raised the trial credit from €5 to €10 — that's 168 hours on an a4 plan or 90 days of a small Postgres instance. Still no card, still no auto-renewal.
Bumped from 14 → 35 days. WAL is now compressed and shipped to two independent storage regions. No price change.